Radar point cloud high-speed preprocessing system and method based on programmable logic

Abstract

The application relates to the technical field of point cloud processing, and discloses a radar point cloud high-speed preprocessing system and method based on programmable logic. The application screens historical distance images with similarity to a current frame meeting preset conditions from a ring-shaped frame buffer, performs interframe motion estimation layer by layer through an image pyramid, generates a motion field and motion confidence in combination with multi-layer iteration correction, greatly improves the calculation accuracy of interframe motion offset, and provides a reliable basis for subsequent motion compensation; adjusts image fusion weights in combination with the reliability of motion confidence and pixel distance values, improves the definition, data stability and detail integrity of the enhanced two-dimensional distance image; performs spatial edge-preserving filtering on the enhanced distance image, restores the two-dimensional image to a three-dimensional point cloud through coordinate inverse transformation, and completes targeted coordinate deviation correction, so that the output radar point cloud provides a high-quality data basis for rear-end target recognition, high-precision imaging and environment perception and the like.

Description

Technical Field

[0001] This application relates to the field of point cloud processing technology, specifically to a high-speed radar point cloud preprocessing system and method based on programmable logic. Background Technology

[0002] As an active detection device, radar is widely used in many fields such as vehicle sensing, security monitoring, terrain mapping and aerial surveying due to its all-weather and all-time operating characteristics. During radar operation, it continuously outputs massive three-dimensional point cloud data streams. The raw point cloud contains invalid information such as ranging noise, clutter interference and motion offset. Direct transmission and application will greatly increase the computational pressure on the back-end parsing, recognition and imaging modules. Therefore, radar point cloud preprocessing has become an indispensable key link in the radar system link.

[0003] The current mainstream radar point cloud preprocessing architecture includes software-based data processing using general-purpose processors and embedded processors, as well as hardware-accelerated processing using programmable logic devices. Among them, programmable logic devices, with their hardware characteristics of parallel computing, pipelined processing, low latency, and high throughput, are better suited to the processing requirements of high-speed continuous data streams in radar and are gradually becoming the mainstream hardware carrier for radar point cloud preprocessing.

[0004] Existing radar point cloud preprocessing solutions based on programmable logic still have several technical shortcomings in practical engineering applications, making it difficult to adapt to the requirements of high-precision and high-real-time radar. Most existing solutions directly call all historical images for inter-frame operations, involving a large number of historical images with large feature differences and no effective motion correlation. In the process of multi-frame image fusion, fixed weights or relying on a single index to adjust the fusion weights are commonly used, resulting in residual noise and blurred details in the fused distance image, and the image enhancement effect does not meet expectations. At the same time, 2D image filtering often adopts a uniform filtering strategy across the entire domain, without correcting for system deviations generated during the projection process and coordinate errors remaining from previous operations. Noise and offset problems in the 2D image stage will be propagated layer by layer to the 3D point cloud, affecting the performance of backend target detection, high-precision imaging and other functions.

[0005] In light of the shortcomings of existing technologies, the technical problem to be solved in this application is how to achieve high-speed online preprocessing of radar point clouds from input to output based on programmable logic devices. Summary of the Invention

[0006] To address the shortcomings of existing technologies, the purpose of this application is to provide a high-speed radar point cloud preprocessing system and method based on programmable logic, which can stably handle the high-frequency continuous point cloud data stream of radar, effectively avoid data congestion, and realize high-speed real-time preprocessing of radar point clouds.

[0007] To achieve the above objectives, this application adopts the following technical solution:

[0008] Firstly, this application provides a high-speed preprocessing method for radar point clouds based on programmable logic, including:

[0009] Within the programmable logic device, the real-time received radar point cloud data stream is mapped online through spherical projection into a two-dimensional distance image of the current frame and stored in a circular frame buffer.

[0010] From the circular frame buffer, select historical distance images whose two-dimensional distance image similarity to the current frame meets the preset conditions, construct image pyramids respectively, and perform inter-frame motion estimation layer by layer to generate motion fields and their motion confidence.

[0011] Motion compensation mapping is performed on historical distance images based on the motion field to obtain compensated historical distance images. The compensated historical distance images are then spatiotemporally fused pixel by pixel with the two-dimensional distance images of the current frame. The fusion weights are adjusted according to the motion confidence and the reliability of the pixel distance values ​​to generate enhanced distance images.

[0012] Spatial edge-preserving filtering is performed on the enhanced range image. The filtered range image is then back-projected into a 3D point cloud through inverse coordinate transformation. After coordinate correction, the preprocessed radar point cloud stream is output.

[0013] Furthermore, the online mapping to the two-dimensional distance image of the current frame includes:

[0014] The radar point cloud data is divided into zones according to the radar detection field of view, and the effective point cloud data is filtered out.

[0015] The spherical projection parameters are dynamically corrected frame by frame. The corresponding projection parameters are configured according to the partition. The effective point cloud data is mapped to two-dimensional pixel coordinates in parallel partitions to obtain the two-dimensional distance image of the current frame.

[0016] The system identifies empty pixels formed by the sparse point cloud within a 2D distance image, performs interpolation processing based on the pixel distribution density of the empty pixel's neighborhood, adds data identifiers to the 2D distance image, and then stores it in a circular frame buffer.

[0017] Furthermore, the partition includes a central main field of view, an edge side field of view, and a blind transition zone;

[0018] The frame-by-frame dynamic correction of spherical projection parameters involves configuring corresponding projection parameters according to the partitions. Specifically, based on the pixel distribution density and angle mapping deviation value of the previous frame's two-dimensional distance image, the spherical projection parameters are iteratively updated frame by frame. Based on the spherical projection parameters, projection curvature and angle compensation coefficients are configured for the three partitions respectively.

[0019] The pixel distance values ​​of the two-dimensional distance image represent the radial distance of the point cloud; when multiple valid point cloud data are mapped to the same pixel, they are combined with echo intensity and effective ranging identifier for hierarchical weighted fusion.

[0020] Furthermore, the generated motion field and its motion confidence include:

[0021] Construct an image pyramid by dividing the selected historical distance image and the current frame's two-dimensional distance image into regions;

[0022] Inter-frame motion estimation is performed layer by layer from the top to the bottom of the pyramid. The displacement vector of the adjacent upper layer is used as the prior offset for the current lower layer calculation, and the motion field is generated by iterative correction.

[0023] By combining matching residuals, regional motion consistency, effective ranging identifiers, and fluctuations in pixel distance values, motion confidence is calculated pixel by pixel, and the motion confidence is corrected by combining image region attributes.

[0024] Furthermore, the aforementioned regional image pyramid construction refers to dividing the image region according to the central main field of view, the edge side field of view, and the blind transition zone, and allocating downsampling ratios to different image regions;

[0025] The layer-by-layer inter-frame motion estimation refers to performing block matching in the top-level image of the pyramid according to image regions to generate an initial motion field, narrowing the pixel matching search range from the top to the bottom of the pyramid, and performing neighborhood smoothing correction.

[0026] The pixel-by-pixel motion confidence is calculated by weighting and fusing the fluctuation range of the matching residual, regional motion consistency, effective ranging identifier, and pixel distance value, and then correcting the motion confidence by configuring correction coefficients according to the image region attributes; the regional motion consistency characterizes the dispersion of the neighborhood displacement vector.

[0027] Further, the generation of the enhanced distance image includes:

[0028] Based on the pixel-by-pixel displacement vector corresponding to the motion field, motion compensation mapping is performed on the historical distance image to obtain the compensated historical distance image;

[0029] The fusion weights are adjusted by combining motion confidence and pixel distance values, and the fusion weights are calibrated by region by combining image region attributes.

[0030] Based on the calibrated fusion weights, the compensated historical distance image and the current frame's two-dimensional distance image are spatiotemporally fused pixel by pixel to generate an enhanced distance image.

[0031] Furthermore, the adjustment of the fusion weights uses motion confidence to characterize the reliability of inter-frame motion matching, integrates the effective ranging identifier and the fluctuation range of pixel distance values ​​to obtain the reliability of pixel distance values, and generates pixel-by-pixel fusion weights by normalizing the motion confidence and the reliability of pixel distance values.

[0032] The regional calibration is to adjust the ratio of fusion weights within the corresponding image region according to the image region attributes.

[0033] Furthermore, the output preprocessed radar point cloud stream includes:

[0034] Spatial edge-preserving filtering is performed on the enhanced distance image according to the partition, and the filtering strategy is dynamically adjusted in combination with the data identifier of the distance image to obtain the filtered distance image;

[0035] By retrieving the projection parameters corresponding to different partitions, and performing inverse coordinate transformation on the filtered distance image partition by partition, the three-dimensional point cloud is restored.

[0036] Data credibility levels are determined based on the reliability of pixel distance values ​​and motion confidence, and coordinate deviation correction is performed on the 3D point cloud according to the data credibility level.

[0037] The corrected 3D point cloud is grouped, cached, and time-ordered, and the preprocessed radar point cloud stream is output.

[0038] Furthermore, the stepwise coordinate inverse transformation refers to reusing the projection curvature and angle compensation coefficients of different partitions during forward spherical projection, mapping the two-dimensional pixel coordinates to three-dimensional coordinates according to a preset coordinate transformation relationship, and compensating for the angle offset caused by forward projection.

[0039] The grouping and timing warping are performed by grouping and storing the original radar point cloud data stream according to the receiving timing sequence. After removing abnormal point cloud data, the receiving timing is aligned in the programmable logic device, and the radar point cloud stream is output.

[0040] Secondly, this application provides a high-speed radar point cloud preprocessing system based on programmable logic, including: a mapping module, an estimation module, a fusion module and an output module;

[0041] The mapping module is used to map the real-time received radar point cloud data stream into a two-dimensional distance image of the current frame through spherical projection within a programmable logic device, and store it in a circular frame buffer.

[0042] The estimation module is used to select historical distance images from the circular frame buffer that have a similarity to the two-dimensional distance image of the current frame that meets the preset conditions, construct image pyramids respectively, and perform inter-frame motion estimation layer by layer to generate motion fields and their motion confidence.

[0043] The fusion module performs motion compensation mapping on the historical distance image based on the motion field to obtain the compensated historical distance image. The compensated historical distance image is then fused with the two-dimensional distance image of the current frame pixel by pixel in a spatiotemporal manner. The fusion weights are adjusted according to the motion confidence and the reliability of the pixel distance values ​​to generate an enhanced distance image.

[0044] The output module is used to perform spatial edge-preserving filtering on the enhanced range image, and then back-project the filtered range image into a 3D point cloud through inverse coordinate transformation. After correcting the coordinates, the preprocessed radar point cloud stream is output.

[0045] Compared with the prior art, the beneficial effects achieved by this application are as follows:

[0046] Historical distance images that meet preset conditions for similarity with the current frame are selected from the circular frame buffer, reducing the amount of computational data of programmable logic devices from the source and effectively shortening the data processing latency. At the same time, inter-frame motion estimation is performed layer by layer through the image pyramid, and motion fields and motion confidence are generated by multi-layer iterative correction, which greatly improves the calculation accuracy of inter-frame motion offset and provides a reliable basis for subsequent motion compensation.

[0047] By combining motion confidence and pixel distance values ​​to adjust image fusion weights, the effective information of multiple frames is fully utilized. While preserving the real-time performance of the current frame image, details are supplemented and random noise is smoothed by using historical images, thereby improving the clarity, data stability, and detail integrity of the enhanced 2D distance image.

[0048] By performing spatial edge-preserving filtering on the enhanced range image, the edge features corresponding to radar detection can be accurately preserved while filtering out image noise. Then, the two-dimensional image is restored to a three-dimensional point cloud through inverse coordinate transformation, and the coordinate deviation is corrected accordingly. This blocks the transmission path of two-dimensional image noise and offset to the three-dimensional point cloud. The output radar point cloud coordinates are more accurate, with less clutter interference, and the spatial shape is more in line with the real detection scene, providing a high-quality data foundation for backend applications such as target recognition, high-precision imaging, and environmental perception. Attached Figure Description

[0049] Figure 1 This is a flowchart of a high-speed radar point cloud preprocessing method based on programmable logic.

[0050] Figure 2 The flowchart shows the logic for mapping the online image to a 2D distance image of the current frame.

[0051] Figure 3 This is a schematic diagram of a high-speed radar point cloud preprocessing system based on programmable logic. Detailed Implementation

[0052] The technical solution of this application will be described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the embodiments and specific features in the embodiments are detailed descriptions of the technical solution of this application, rather than limitations thereof.

[0053] Example 1

[0054] like Figure 1 As shown, this embodiment provides a high-speed preprocessing method for radar point clouds based on programmable logic, including: S1, within the programmable logic device, the real-time received radar point cloud data stream is mapped online into a two-dimensional distance image of the current frame through spherical projection, and stored in a circular frame buffer.

[0055] Specifically, such as Figure 2 As shown, the online mapping to the two-dimensional distance image of the current frame includes:

[0056] Within a programmable logic device, radar point cloud data is partitioned and effective point cloud data is filtered according to the radar detection field of view.

[0057] The spherical projection parameters are dynamically corrected frame by frame. The corresponding projection parameters are configured according to the partition. The effective point cloud data is mapped to two-dimensional pixel coordinates in parallel partitions to obtain the two-dimensional distance image of the current frame.

[0058] The system identifies empty pixels formed by the sparse point cloud within a 2D distance image, performs interpolation processing based on the pixel distribution density of the empty pixel's neighborhood, adds data identifiers to the 2D distance image, and then stores it in a circular frame buffer.

[0059] The zones include the central main field of view, the edge side fields of view, and the blind transition zone;

[0060] The spherical projection parameters are dynamically corrected frame by frame, and the corresponding projection parameters are configured according to the partition. Specifically, the spherical projection parameters are iteratively updated frame by frame based on the pixel distribution density and angle mapping deviation value of the previous frame's two-dimensional distance image, and the projection curvature and angle compensation coefficients are configured for the three partitions respectively based on the spherical projection parameters.

[0061] The pixel distance value of the two-dimensional distance image represents the radial distance of the point cloud; when multiple valid point cloud data are mapped to the same pixel, the echo intensity and the effective ranging identifier are combined for hierarchical weighted fusion.

[0062] The spatial distribution of point clouds, signal-to-noise ratio, and geometric distortion characteristics in different regions of the radar detection field of view have inherent differences. Performing projection calculations across the entire field of view will result in a decrease in projection matching degree and ineffective consumption of hardware computing resources of programmable logic devices. Radar point cloud data streams are generally mixed with out-of-range data, environmental stray echoes, and angle jump data. Such data has no effective detection value. If it is directly entered into the projection link, it will occupy the device's computing channels and on-chip storage resources, generate abnormal pixels, and interfere with the quality of subsequent image generation.

[0063] The programmable logic device continuously captures serially input radar point cloud data packets. Each data packet is embedded with fields such as three-dimensional spatial coordinates, detection angle, radial distance, and echo intensity. The device has a preset angle discrimination logic that analyzes the detection angle corresponding to a single set of point clouds point by point, dividing the radar detection field of view into three continuous and non-overlapping regions: the central main field of view, the edge side field of view, and the blind zone transition zone. Here, the preset angle configuration is that the radar's global horizontal field of view is set to -90° to +90°, of which the range of -30° to +30° is defined as the central main field of view, the ranges of -90° to -30° and +30° to +90° are defined as the edge side field of view, and the range of ±8° on both sides of the field of view boundary is defined as the blind zone transition zone. Those skilled in the art can adjust the preset angle in actual operation.

[0064] After spatial domain partitioning and aggregation, the point cloud data enters the invalid data filtering process. First, range filtering is performed. The device's preset radar rated radial detection range is 0.5m to 200m. Point clouds with radial distances less than 0.5m or greater than 200m are marked as invalid data and physically isolated. Then, angle distortion filtering is performed. When the difference in detection angle between consecutive adjacent point clouds exceeds 0.8°, it is determined to be angle jump clutter and is removed. Those skilled in the art can adjust the filtering conditions accordingly in the actual process. Valid point cloud data is output, and the marked invalid data will no longer participate in any subsequent calculations.

[0065] This application achieves centralized aggregation of point clouds with similar characteristics, reduces the data volume of actual computation of programmable logic devices, ensures the real-time processing capability of continuous data streams, and blocks abnormal data from flowing into the projection stage, maintaining the basic quality of image generation; the effective point cloud data is the data source for subsequent spherical projection parameter correction and partitioned parallel coordinate mapping, ensuring that subsequent projection parameters accurately match the corresponding field of view.

[0066] In actual radar operations, the spatial position of the detected target and environmental electromagnetic interference are in a dynamic state. Fixed spherical projection parameters will cause the inter-frame angle mapping deviation to accumulate continuously, which will lead to distortion problems such as overall offset and geometric stretching in the two-dimensional distance image. The geometric distortion patterns of the central main field of view, the edge side field of view, and the blind transition zone are significantly different, and a single projection parameter cannot meet the full field of view mapping requirements. In radar scanning operations, there are also scenarios where multiple sets of three-dimensional point clouds are mapped to the same two-dimensional pixel. Simply using single data assignment or simple averaging calculation cannot distinguish the signal reliability of different point clouds, which will directly reduce the quantization accuracy of pixel distance values.

[0067] Among them, frame-by-frame dynamic correction of spherical projection parameters refers to iteratively updating the global spherical projection parameters using the statistical features of the previous frame's two-dimensional distance image; partition configuration of corresponding projection parameters refers to configuring projection curvature and angle compensation coefficients for the three field-of-view partitions based on the updated global spherical projection parameters; partition parallel mapping refers to enabling independent computing channels of programmable logic devices to synchronously complete the conversion of three-dimensional coordinates of each partition to two-dimensional pixel coordinates; and hierarchical weighted fusion refers to combining echo intensity and effective ranging identifiers to complete the final pixel assignment for multi-point cloud data of the same pixel.

[0068] After each frame of image is generated, the programmable logic device retrieves the two-dimensional distance image of the previous frame from the circular frame buffer, traverses the entire image to count the pixel distribution density and angle mapping deviation value, and keeps the parameter update cycle synchronized with the point cloud frame rate. Here, the point cloud frame rate is set to 50 frames / second to ensure that the parameters adapt to changes in the field conditions in real time. The pixel distribution density D is defined as the ratio of the total number of effective pixels in the image to the total number of pixels in the image, and the angle mapping deviation value Δθ is the average deviation between the theoretical mapping angle and the actual pixel corresponding angle.

[0069] The fundamental parameters of the global spherical projection, namely P, are updated using an incremental iterative algorithm. new =P old +λ1×D-λ2×Δθ, where P old P represents the spherical projection parameters of the previous frame. new To update the spherical projection parameters, λ1 and λ2 are preset iteration step size coefficients, where λ1 is set to 0.01 to control the response sensitivity of pixel distribution density to projection parameter correction, and λ2 is set to 0.02 to control the response sensitivity of angle mapping deviation to projection parameter correction. Those skilled in the art can adaptively adjust these coefficients within the range of 0.005 to 0.05 according to the radar system's frame rate, point cloud density, and scene dynamics; for example, if the previous frame parameter P... old =0.98, statistically we obtain D=0.92 and Δθ=0.03rad, substituting them into the calculation to get P new=0.98+0.01×0.92-0.02×0.03=0.9886, completing the single-frame parameter iterative update.

[0070] After the global spherical projection parameters are updated, differentiated compensation parameters are loaded according to the partition type. Specifically, the projection curvature of the central main field of view is set to 1, and the angle compensation coefficient is set to 0.02 rad to adapt to the low-distortion imaging characteristics of this region; the projection curvature of the edge side field of view is set to 1.15, and the angle compensation coefficient is set to 0.18 rad to compensate for the stretching distortion caused by the large field of view; the projection curvature of the blind transition zone is set to 1.08, and the angle compensation coefficient is set to 0.3 rad to achieve a smooth imaging transition between the effective field of view and the detection blind zone.

[0071] The projection curvature reflects the degree of nonlinear compression of different viewing angle regions by spherical projection. The central main viewing field has a small viewing angle and slight geometric distortion, so the projection curvature is set to the baseline value of 1. The edge side viewing field has a large viewing angle, which causes significant image stretching and distortion, so the projection curvature is set to 1.15 to increase compensation for projection nonlinearity. The blind transition area is between the effective viewing field and the blind area, with a moderate degree of distortion, so the projection curvature is set to 1.08.

[0072] The angle compensation coefficient is used to compensate for the angle offset introduced by the spherical projection in each zone. Its value increases with the degree of the field of view edge. The minimum offset in the central region is 0.02 rad, the offset in the edge region is larger and is 0.18 rad, and the offset in the transition zone is 0.3 rad to ensure smooth boundary connection. Those skilled in the art can make adaptive adjustments to the above spherical projection parameters according to the specific radar optical system parameters and the measured distortion calibration results.

[0073] The system synchronously reads valid point cloud data from the corresponding partitions and performs spherical projection coordinate transformation based on global spherical projection parameters and partition projection coefficients. This transforms the three-dimensional coordinates of the point cloud's azimuth, elevation, and radial distance into row and column pixel coordinates of a two-dimensional image. The radial distance of the point cloud is then assigned as a pixel value according to its coordinate position. This pixel distance value uniformly represents the radial detection distance of the original point cloud. Let the three-dimensional parameters of the point cloud be azimuth Az, elevation E1, and radial distance R, the partition projection curvature be k, the angle compensation coefficient be Δα, and the total number of columns in the image be U. t Total number of rows V t The quantitative conversion formula for two-dimensional pixel column coordinates u and row coordinates v is u=U. t / 2×((Az+Δα) / (k×π / 2)+1), v=V t / 2×((El+Δα) / (k×π / 4)+1); Taking the central main field of view as an example, take U t =640、V t=480, k=1, Δα=0.02rad, single point cloud Az=0rad, El=0rad, R=60m, substituting into the calculation, we get u=640 / 2×((0+0.02) / (1×π / 2)+1)=320×(0.02 / 1.5708+1)≈324, v=480 / 2×((0+0.02) / (1×π / 4)+1)=240×(0.02 / 0.7854+1)≈246, the pixel coordinates are (324,246), and the pixel position is assigned a pixel distance value of 60m.

[0074] When two or more sets of valid point cloud data are detected to be associated with the same row and column coordinates within a temporary pixel matrix, a hierarchical weighted fusion logic is initiated. First, the hierarchy is determined based on the effective ranging identifier inherent in each set of point clouds. Point clouds identified as having effective ranging are assigned to a high-confidence hierarchy, while those identified as having questionable ranging are assigned to a low-confidence hierarchy. After the hierarchy is determined, weights are allocated within the same hierarchy based on echo intensity; the higher the echo intensity, the greater the corresponding weight. The sum of all data weights within a single hierarchy is normalized. After weighted summation, the final value is written to the pixel location. The echo intensity of a single set of point clouds is defined as S. i The total echo intensity within the layer is ∑S i Single-group point cloud normalized weight w i =S i / ∑S i Pixel distance value Val = ∑(w i ×R i ), R i This represents the radial distance of a single point cloud set; for example, if three high-confidence level point clouds are superimposed on the same pixel, with radial distances of 58m, 60m, and 62m respectively, the echo intensities are 80, 100, and 70 respectively; the total echo intensity ∑S i =250, corresponding to normalized weights of 0.32, 0.4, and 0.28 respectively; the weighted calculation yields Val = 0.32×58 + 0.4×60 + 0.28×62 = 59.92m, which is used as the pixel distance value; after all partition mapping and pixel fusion operations are completed, the pixels of each partition are stitched together to form a single-frame two-dimensional distance image.

[0075] This application continuously offsets the mapping deviation caused by scene changes and inherent system errors, maintains the consistency and stability of inter-frame imaging, specifically compensates for geometric distortion in different fields of view, and improves the geometric accuracy of two-dimensional distance images from the root of projection; it fully utilizes the parallel computing power of programmable logic devices to match the real-time processing requirements of high frame rate radar data streams; it integrates the effective information of multiple sets of detection data to avoid low-reliability data dominating pixel values ​​and improve the quantization accuracy of pixel distance values.

[0076] Radar scanning mechanisms have inherent sampling intervals. In long-range detection areas, field of view edges, and blind transition zones, sparse three-dimensional point cloud space is prone to occur. This physical feature is mapped to a two-dimensional distance image as empty pixels with no effective values. Empty pixels cause interruptions in image data continuity, which not only destroys the integrity of a single frame image but also leads to computational anomalies in subsequent steps such as inter-frame similarity determination, motion estimation, and image fusion. An empty pixel refers to a pixel in a two-dimensional distance image that is not mapped by any effective point cloud and has no effective radial distance. The pixel distribution density of the neighborhood refers to the ratio of the number of effective pixels in a fixed neighborhood window centered on an empty pixel to the total number of pixels in the window.

[0077] The programmable logic device initiates a pixel-by-pixel traversal, scanning the two-dimensional distance image row by row in row priority order. It verifies the existence of a valid pixel distance value at each pixel position, simultaneously marking the row and column coordinates of all empty pixels during the traversal. For each marked empty pixel, a 5×5 pixel neighborhood window is defined, the number of valid pixels within the window is counted, and the pixel distribution density of the neighborhood is calculated. Based on the pixel distribution density, an interpolation strategy is determined: when the pixel distribution density is greater than 0.6, the neighborhood data is considered sufficient, and weighted interpolation of nearby neighbor pixels is used; when the pixel distribution density is between 0.3 and 0.6, the neighborhood data is considered moderate, and linear interpolation with a full-area window is used; when the pixel distribution density is less than 0.3, the neighborhood data is considered extremely sparse, and to avoid interpolation distortion, the original state of the empty pixel is preserved, and no interpolation operation is performed.

[0078] This application comprehensively identifies empty pixels caused by sparse point clouds, ensuring that no abnormal points are missed. While filling in the gaps in image data, it avoids image distortion caused by blind interpolation in sparse neighborhoods, balancing image integrity and the authenticity of the original detection data. It also significantly reduces the interference of empty pixels on subsequent operations.

[0079] The circular frame buffer inside the programmable logic device is responsible for the cyclic storage and fast retrieval of multiple consecutive frames of images. Unmarked original images cannot achieve full-process data traceability and accurate partition retrieval. Data identification refers to the set of status markers configured by the programmable logic device for two-dimensional distance images and internal pixels, including multiple types of markers such as image partition markers, pixel type markers, and data validity markers. The circular frame buffer refers to the circular on-chip storage unit built inside the device, which is used to continuously cache multiple frames of historical distance images.

[0080] The system iterates through the 2D distance image of the current frame, adds partition identifiers to the image based on the previous partitioning results, and marks the pixel type pixel by pixel, distinguishing between two types of pixel states: original mapped valid pixels and interpolated completed pixels. All identifier fields are bound to the pixel distance values ​​and stored to form an image data packet carrying data tags. The circular frame buffer has a preset fixed storage depth, which is set to 16 frames here, meaning that the buffer can cache 16 consecutive frames of 2D distance images at the same time. The storage rule adopts a first-in-first-out mode. New frame image data packets are written to the end of the buffer according to the frame reception time. When the number of stored frames reaches the maximum depth, the earliest historical frame stored at the head of the buffer is automatically removed.

[0081] This application enables data traceability throughout the entire lifecycle of images and pixels, improving the execution efficiency and discrimination accuracy of multi-frame linkage operations; the circular frame buffer enables continuous caching of multiple historical images under limited hardware storage resources, adapting to the processing scenario of continuous radar data streams, and the first-in-first-out storage rule ensures the rationality of the timing of historical image retrieval, guaranteeing the closed-loop flow of the entire radar point cloud preprocessing data stream.

[0082] S2. From the circular frame buffer, select historical distance images whose two-dimensional distance image similarity to the current frame meets the preset conditions, construct image pyramids respectively, and perform inter-frame motion estimation layer by layer to generate motion fields and their motion confidence.

[0083] Specifically, generating the sports field and its motion confidence includes:

[0084] Select historical distance images from the circular frame buffer that have a similarity to the current frame's two-dimensional distance image that meets preset conditions;

[0085] Construct an image pyramid by dividing the selected historical distance image and the current frame's two-dimensional distance image into regions;

[0086] Inter-frame motion estimation is performed layer by layer from the top to the bottom of the pyramid. The displacement vector of the adjacent upper layer is used as the prior offset for the current lower layer calculation, and the motion field is generated by iterative correction.

[0087] By combining matching residuals, regional motion consistency, effective ranging identifiers, and fluctuations in pixel distance values, motion confidence is calculated pixel by pixel, and the motion confidence is corrected by combining image region attributes.

[0088] Among them, constructing an image pyramid by region refers to dividing the image region into the central main field of view, the edge side field of view, and the blind transition zone, and assigning downsampling rates to different image regions;

[0089] Layer-by-layer inter-frame motion estimation refers to performing block matching by image region in the top layer of the pyramid to generate an initial motion field, narrowing the pixel matching search range from the top layer of the pyramid to the bottom layer, and performing neighborhood smoothing correction.

[0090] Pixel-by-pixel motion confidence is obtained by weighted fusion of matching residuals, regional motion consistency, effective ranging identifiers, and fluctuations in pixel distance values. Correction coefficients are configured based on image region attributes to adjust the motion confidence. Regional motion consistency characterizes the dispersion of neighborhood displacement vectors.

[0091] A circular frame buffer continuously stores multiple frames of 2D distance images in a sequential manner. If all historical frames are used for motion estimation, the computation channels and storage resources of the programmable logic device will be continuously occupied, increasing the overall processing latency and failing to meet the design requirements of high-speed preprocessing. Historical images with large differences in features between frames do not have an effective motion correlation with the current frame. Using them in the calculation will introduce matching noise and reduce the accuracy of motion estimation.

[0092] Selecting historical distance images that meet the preset similarity conditions involves retrieving all historical distance images cached in the circular frame buffer. This means determining similarity through feature comparison, filtering based on a preset threshold, and identifying the objects to participate in subsequent calculations. Following a first-in-first-out (FIFO) order, each frame of historical distance images in the circular frame buffer is sequentially read from the current frame's two-dimensional distance image, and pixel distribution features are extracted. The overlap of pixel distribution features is used as the criterion, with an overlap threshold set at 75%. Those skilled in the art can adjust the threshold accordingly in practice. The feature overlap between a single historical distance image and the current frame is calculated frame by frame. The calculation process is implemented by pixel traversal, and the final overlap is obtained by statistically analyzing the percentage of pixels in the same feature region.

[0093] During the filtering process, the frame number, feature overlap, and filtering results of each historical distance image are recorded synchronously. Historical distance images with a feature overlap higher than 75% are marked as valid images and proceed to the next processing stage. Historical distance images with a feature overlap lower than 75% are marked as invalid images, retained in the circular frame buffer, and do not participate in subsequent calculations. After filtering is completed, the valid historical distance images and the current frame form an image pairing group.

[0094] This application reduces the scale of computational data from the source, reduces hardware resource consumption and processing latency, and eliminates computational interference from unrelated images, ensuring the quality of basic data for motion estimation, stably supporting subsequent image pyramid construction operations, and avoiding invalid data causing abnormalities in subsequent module operations.

[0095] Radar 2D range images have high original resolution. Directly performing global motion estimation would significantly increase the computational load of block matching. There are significant differences in point cloud density, imaging distortion, and motion change patterns in different regions. A uniform downsampling rule would cause the loss of local motion features. The regional image pyramid construction refers to dividing the image region according to the central main field of view, the edge side field of view, and the blind transition zone. A downsampling factor is set for each region, and a multi-level image pyramid is generated by downsampling step by step.

[0096] Image data from the image pairing group is read, and a single frame image is divided into three regions based on the image's built-in partition identifier. Each of the three regions is connected to an independent parallel downsampling operation channel within the device. A fixed downsampling factor is configured for each of the three regions to characterize the compression ratio of image rows and columns of pixels during each downsampling stage. The downsampling factor is set to 2 for the central main field of view, 3 for the edge side field of view, and 4 for the blind transition area. This is because the central main field of view has a high point cloud density and good imaging quality, so a lower downsampling factor is used to retain more motion details for accurate matching. The edge side field of view and the blind transition area have lower point cloud density and increased noise, so a higher downsampling factor is used to reduce noise interference to block matching and reduce computational load. Those skilled in the art can adjust the downsampling factor within the range of 2 to 6 according to the computational resource capacity and real-time requirements of the programmable logic device.

[0097] The pyramid employs a multi-level structure design, uniformly setting the total number of pyramid layers to 4. The original resolution image serves as the bottom layer of the pyramid, and downsampling is performed layer by layer upwards. This is because too few layers would result in excessively high resolution at the top layer and an excessively large search range, increasing computation time; while too many layers would result in insufficient information per layer, leading to a decrease in motion estimation accuracy. Those skilled in the art can adjust the number of layers to within the range of 3 to 5 based on the actual image resolution. The downsampling operation for a single region and a single layer is implemented using the neighborhood mean sampling method, taking the average value of all original pixels in the neighborhood of the target pixel as the value of the new pixel after downsampling. After each layer of downsampling is completed, a partition image and layer parameters for that layer are generated. The layer parameters include the current layer, partition type, image resolution, and downsampling ratio. After completing the operation layer by layer from the bottom to the top, the image data of all layers and all partitions are integrated to form the current frame image pyramid and the historical frame image pyramid, respectively.

[0098] This application adapts to the imaging characteristics of different fields of view, retains the core motion features of each region, and uses a multi-level pyramid structure to achieve coarse-to-fine feature analysis. The parallel computing architecture fully leverages the hardware advantages of programmable logic devices to improve the overall processing speed. It provides a standardized computing carrier for subsequent top-down layer-by-layer execution of inter-frame motion estimation, and the partitioned and isolated data format can also ensure that the block matching operation accurately matches the regional motion patterns.

[0099] The top layer of the image pyramid has the lowest resolution and focuses on representing the global motion trend, while the bottom layer has the highest resolution and focuses on representing the local motion details. Unrestricted pixel matching increases computation time, and simple block matching is also prone to noise from abrupt changes in displacement vectors. Layer-by-layer inter-frame motion estimation refers to performing block matching by partition at the top layer of the pyramid to generate the initial motion field. When calculating layer by layer downwards, the displacement vector of the previous layer is mapped to the prior offset of the current layer, the matching search range is narrowed simultaneously, and neighborhood smoothing correction is performed. After multiple iterations, the motion field is obtained.

[0100] The device starts its computation from the fourth layer of the pyramid, the top layer. It first divides the top layer image into blocks, with each block having a pixel size of 8×8. Inter-frame block matching is performed using these blocks as the basic unit. Block matching employs a pixel difference summation method, traversing all positions within the search range and selecting the position with the smallest total pixel difference as the optimal matching position. The displacement vector of the corresponding image block is calculated, and the combination of all image block displacement vectors forms the initial motion field of the top layer. When moving to the next layer, the third layer, the displacement vectors from the previous layer are scaled according to the layer ratio, converting them into prior offsets for the current layer's pixels. For each layer below, the pixel matching search range is reduced by 20% compared to the previous layer. A finite search interval is defined centered on the prior offset, and precise block matching is performed within this interval, generating the current layer's displacement vector. The 20% reduction in search range is based on an empirical value derived from the resolution multiplication relationship between pyramid layers, aiming to narrow the search range layer by layer while retaining sufficient matching margin to cope with the progressive refinement of motion vectors. This ratio can be adjusted within the range of 15% to 30% depending on the scene's motion amplitude.

[0101] After the displacement vector calculation of each layer is completed, the neighborhood smoothing correction operation is started. A 3×3 neighborhood window is defined with a single pixel as the center. The mean value of all displacement vectors in the window is taken to replace the original vector of the current pixel to suppress vector mutation noise. The operation of the second layer and the first layer, which is the bottom layer, is completed in sequence according to this logic. After the bottom layer operation is completed, the displacement vector of all pixels in the entire image is obtained, that is, the motion field.

[0102] This application reduces the number of invalid matches, narrows the search range layer by layer to improve computational efficiency, and adapts to high-speed processing requirements; neighborhood smoothing correction ensures the spatial continuity of displacement vectors, which fits the actual motion state of radar-detected targets; the motion field represents the positional offset relationship of all pixels between frames, which is the basis for subsequent motion compensation mapping of historical distance images, and the accuracy and continuity of the vectors determine the final effect of motion compensation.

[0103] Motion confidence is used to quantify the reliability of motion estimation results. A single index cannot comprehensively evaluate the effectiveness of displacement vectors. Matching bias, local motion stability, and the quality of the original ranging data all affect the confidence determination result. Different image partitions have different noise intensities and data fluctuation characteristics, and uniform calculation rules will cause partition evaluation bias. Pixel-by-pixel motion confidence calculation refers to fusing the fluctuation range of matching residuals, regional motion consistency, effective ranging identifiers, and pixel distance values, and obtaining the motion confidence through weighted fusion. Among them, regional motion consistency characterizes the dispersion of neighborhood displacement vectors. Then, according to the image partition attributes, a correction coefficient is configured to perform partition correction on the motion confidence to obtain the final motion confidence.

[0104] The programmable logic device traverses the motion field pixel by pixel, extracting four indicators sequentially. The matching residual is the numerical difference between corresponding pixels in frames after the current pixel is matched. The regional motion consistency is quantified by calculating the variance of the displacement vector in the 3×3 neighborhood. The larger the variance, the higher the dispersion of the neighborhood vector and the worse the motion consistency. The effective ranging indicator is a binary marker, with a value of 1 for a valid marker and 0 for an invalid marker. The fluctuation range of the pixel distance value is quantized by taking the standard deviation of the distance values ​​of the current pixel in three consecutive frames. The four indicators are mapped to the [0,1] interval through the min-max normalization method to eliminate the difference in dimensions. The device internally presets the weight coefficients of the four indicators, namely, the matching residual weight 0.3, the regional motion consistency weight 0.25, the effective ranging indicator weight 0.25, and the pixel distance value fluctuation range weight 0.2, with the total weight being 1. Those skilled in the art can adjust the weights accordingly in the actual process.

[0105] Among them, the matching residual directly reflects the accuracy of inter-frame pixel matching and is the most direct indicator of motion estimation reliability. The regional motion consistency reflects the coordination of neighborhood motion and is used to detect whether the matching is affected by abnormal vector interference. The effective ranging identifier directly indicates the credibility of the original data. The fluctuation amplitude of the pixel distance value reflects the temporal stability of the original ranging and serves as an auxiliary correction indicator. The above weights are determined based on the prior assessment of the contribution of each indicator to the reliability of motion estimation. Those skilled in the art can make adaptive adjustments according to the noise characteristics of the specific radar system and the dynamic characteristics of the target scene, under the constraint that each weight is not negative and the sum is 1.

[0106] Motion confidence is calculated using a weighted summation formula. The quantified values ​​of the four indicators are multiplied by their corresponding weights and then summed to obtain the motion confidence of a single pixel. After the initial calculation, differential correction coefficients are applied according to the pixel's region: 0.95 for the central main field of view, 0.85 for the edge side field of view, and 0.7 for the blind transition zone. The central main field of view has high data quality and reliable motion estimation results, so its correction coefficient is close to 1 to retain the original confidence. The edge side field of view has increased geometric distortion and decreased matching accuracy, so its correction coefficient is reduced to 0.85 to appropriately lower the confidence. The blind transition zone has the worst data continuity and the strongest noise, so its correction coefficient is reduced to 0.7 to significantly lower the confidence, avoiding excessive influence of low-quality motion estimation results on subsequent fusion. The correction coefficient ranges from 0 to 1, and those skilled in the art can adjust it based on the measured data quality of each region. The initial motion confidence of a single pixel is multiplied by the corresponding region correction coefficient to obtain the final motion confidence. After all pixels are calculated, a motion confidence corresponding to the motion field is formed.

[0107] This application performs a comprehensive evaluation from multiple dimensions, including matching accuracy, motion stability, and reliability of the original data, resulting in a more objective and comprehensive confidence score calculation. It adapts to the noise and data characteristics of different fields of view, offsets the evaluation error caused by inherent interference in the partitions, and further improves the rationality of the confidence score. The motion confidence score provides an adaptive adjustment basis for the subsequent image spatiotemporal fusion process.

[0108] S3. Motion compensation mapping is performed on the historical distance image based on the motion field to obtain the compensated historical distance image. The compensated historical distance image is then spatiotemporally fused pixel by pixel with the two-dimensional distance image of the current frame. The fusion weight is adjusted according to the motion confidence and the reliability of the pixel distance value to generate the enhanced distance image.

[0109] Specifically, generating the enhanced distance image includes:

[0110] Based on the pixel-by-pixel displacement vector corresponding to the motion field, motion compensation mapping is performed on the historical distance image to obtain the compensated historical distance image;

[0111] The fusion weights are adjusted by combining motion confidence and pixel distance values, and the fusion weights are calibrated by region by combining image region attributes.

[0112] Based on the calibrated fusion weights, the compensated historical distance image and the current frame's two-dimensional distance image are spatiotemporally fused pixel by pixel to generate an enhanced distance image.

[0113] Among them, adjusting the fusion weight is to use motion confidence to characterize the credibility of inter-frame motion matching, integrate the effective ranging identifier and the fluctuation range of pixel distance value to obtain the reliability of pixel distance value, and then normalize the motion confidence and the reliability of pixel distance value to generate pixel-by-pixel fusion weights.

[0114] Regional calibration adjusts the fusion weight ratio within a corresponding image region according to the image region attributes.

[0115] During radar detection, there is relative motion between the radar body and the target. Pixels with the same name in different frames of two-dimensional distance images will have positional shifts. Directly fusing the original historical images with the current frame image will result in problems such as ghosting, blurred edges, and data misalignment, and the supplementary information value of the historical images cannot be utilized. The displacement vector is the two-dimensional offset corresponding to the pixel coordinates within the motion field, representing the pixel offset in the row and column directions of the image, respectively. Motion compensation mapping is a coordinate transformation process that remaps each pixel of the historical distance image to the registration coordinate position according to the offset rules of the displacement vector.

[0116] First, the historical distance image and the displacement vector of the motion field are loaded synchronously, and the row and column dimensions of the two sets of data are completely consistent. The programmable logic device uses pixels as the smallest unit of operation and traverses the entire historical distance image in row priority order. For each original coordinate point, the row offset vector and column offset vector corresponding to that position are extracted. For example, the single pixel row offset value range is set to -12 to +12 pixels, and the column offset value range is set to -10 to +10 pixels. Positive and negative values ​​represent the offset direction in the two dimensions, respectively. The maximum pixel offset between frames is estimated based on the maximum angular velocity of the radar system and the inter-frame time interval. The offset range in the row direction is slightly larger than that in the column direction because the offset generated by the radar scanning motion in the row direction (corresponding to the azimuth direction) is usually greater than that in the column direction (corresponding to the elevation direction). Those skilled in the art can adjust the offset range between ±5 and ±20 pixels according to the motion characteristics and frame rate of the radar system. Combining the original pixel coordinates and the offset vector, the target coordinates of the pixel after compensation are calculated. During the calculation, the original row coordinates, original column coordinates, row offset, column offset, target row coordinates, and target column coordinates of the pixel coordinates are recorded synchronously.

[0117] During the traversal, boundary judgment logic is set. If the calculated target coordinates exceed the valid row and column range of the image, the pixel is judged as an invalid pixel that is out of bounds and the data of that group is directly discarded and will not participate in subsequent image stitching. For pixels whose target coordinates are within the valid range, the original pixel distance value is temporarily stored in the image buffer area corresponding to the target coordinates. After all pixels have completed coordinate mapping and validity judgment, the pixel data in the image buffer area is integrated. For points where multiple original pixels are mapped to the same target coordinates, the average pixel distance value is used to complete the merging and assignment, generating a historical distance image aligned with the coordinates of the current frame image.

[0118] This application achieves precise coordinate mapping, eliminates positional deviations caused by relative motion between frames, and realizes spatial registration between historical and current images; the spatial position of the compensated historical distance image matches the current frame, and the data format and pixel dimension are unified, providing a standard paired data source for subsequent processing.

[0119] Using fixed fusion weights cannot distinguish between the motion matching quality and ranging data quality of different pixels, resulting in insufficient high-confidence data and excessive superposition of low-confidence data. At the same time, there are significant differences in noise levels and data stability in the central main field of view, edge side fields of view, and blind transition areas, and weights generated by unified rules cannot be adapted to the partitioned conditions. Among them, the reliability of pixel distance values ​​is used to characterize the overall credibility of single-pixel ranging data, which is obtained by integrating the effective ranging identifier and the fluctuation range of pixel distance values. Motion confidence characterizes the credibility of inter-frame motion matching results. Normalization is used to eliminate the dimensional differences of different indicators. Regional calibration is to adjust the ratio coefficient of fusion weights according to the imaging characteristics of different field of view partitions.

[0120] The programmable logic device performs operations in stages. The first stage completes the reliability calculation of the pixel distance value, which involves retrieving the effective ranging identifier corresponding to the pixel and the fluctuation range of the pixel distance value. The effective ranging identifier adopts a binary quantization rule, assigning a value of 1 when the identifier is valid ranging and 0 when the identifier is invalid ranging. The fluctuation range of the pixel distance value is quantized using the standard deviation of the pixel distance values ​​at the same position in three consecutive frames, with a value range of 0 to 8. After normalizing the fluctuation range to the [0,1] range, the inverse index is taken to obtain the stability of the pixel distance value, i.e., stability = 1 - fluctuation range / 8. The higher the stability, the more stable and reliable the ranging data. The integrated weight of the two indicators is preset, configured as an integrated weight of 0.6 for the effective ranging identifier and an integrated weight of 0.4 for the stability of the pixel distance value. The reliability of the single pixel distance value is calculated by weighted summation, i.e., reliability = 0.6 × effective ranging identifier + 0.4 × (1 - fluctuation range / 8).

[0121] For example, if a pixel's effective ranging indicator is 1 and its fluctuation range is 2, then the reliability is 0.6×1+0.4×(1-2 / 8)=0.6+0.3=0.9; if another pixel's effective ranging indicator is 1 and its fluctuation range is 7, then the reliability is 0.6×1+0.4×(1-7 / 8)=0.6+0.05=0.65. The larger the fluctuation range, the lower the reliability.

[0122] The second stage involves data normalization and basic weight generation. The device reads the motion confidence and pixel distance reliability values ​​respectively. Since the original value ranges of these two types of data differ, a minimum-maximum normalization method is used to map both sets of data to the 0-1 range. Normalization is performed column-by-column, that is, the minimum and maximum values ​​of motion confidence and pixel distance reliability are calculated column by column, and the normalized values ​​are calculated using the above method. After traversal, two sets of normalized data are obtained. The device presets a basic matching coefficient, with a basic ratio of 0.55 for motion confidence and 0.45 for pixel distance reliability. The sum of these two ratios is always 1. This is because motion confidence reflects the accuracy of inter-frame motion matching and has a more direct impact on the registration quality of the fused image. Pixel distance reliability reflects the inherent quality of single-frame ranging data and serves as an auxiliary adjustment factor; those skilled in the art can adjust the ratios accordingly in practice. Pixel-by-pixel, the two sets of normalized values ​​and the matching coefficient are combined according to W=α×C. norm +β×R norm The fusion weights are calculated, where W is the fusion weight, α=0.55 is the base proportion of motion confidence, β=0.45 is the base proportion of pixel distance value reliability, and C... norm R is the normalized motion confidence score. norm The reliability of the normalized pixel distance value.

[0123] The third stage completes the regional weight calibration. Based on the image partitioning identifier, the device divides the fusion weights into the central main field of view, the edge side field of view, and the blind transition zone. Correction coefficients are configured for different partitions. The central main field of view has stable data and low noise, so a small correction coefficient is set to fine-tune the proportion of the two types of indicators. The edge side field of view has increased distortion and interference, so the weight of motion confidence is increased. The blind transition zone has poor data continuity and large ranging fluctuations, so the weight of pixel distance value reliability is increased. The correction coefficients of each partition act on the proportion of motion confidence and pixel distance value reliability in the basic ratio through multiplicative factors.

[0124] For example, the correction coefficient for the central main field of view is 0.98 to 1.02, indicating that the fine-tuning of the basic ratio does not exceed ±2%, as the two indicators in this area are already relatively reliable. The correction coefficient for the edge side field of view is 0.9 to 1.1, indicating that the proportion of motion confidence can be adjusted within the range of 90% to 110% of the basic value. Because the increased distortion in this area leads to a decrease in motion matching reliability, the relative proportion of pixel distance value reliability needs to be appropriately increased to compensate. The correction coefficient for the blind transition area is 0.85 to 1.15, indicating that the adjustment range can reach ±15%. Because the data continuity in this area is poor, the proportion of pixel distance value reliability needs to be increased to a greater extent to suppress the error propagation caused by unreliable motion estimation. The specific values ​​of the correction coefficients need to be calibrated according to the measured noise level and data quality of each zone of each radar system. The internal ratio of the fusion weight is adjusted pixel by pixel according to the zone correction. After all zones have been calculated, the calibrated fusion weights are generated.

[0125] This application quantifies the reliability of pixel ranging, generates basic weights by combining motion confidence, and ensures that the weight allocation fits the true state of the data; it adapts to the imaging characteristics of different fields of view, making up for the adaptation defects of the unified basic weights in complex areas; the calibrated fusion weights correspond to the pixels of the two images to be fused, and adaptively matches the data quality and regional attributes of each point, providing a precise weighting basis for pixel-by-pixel spatiotemporal fusion.

[0126] The compensated historical image can supplement the missing details and smooth random noise in the current frame image, while the current frame image can ensure the real-time nature of the detection information. The combination of the two can achieve image enhancement, and the fusion effect is entirely determined by the fusion weight. Among them, pixel-by-pixel spatiotemporal fusion refers to the weighted calculation of the pixel value of the compensated historical distance image and the pixel value of the current frame two-dimensional distance image at the same coordinate position, combined with the corresponding fusion weight, to fuse the historical data and real-time data in the temporal dimension into a single pixel value.

[0127] The programmable logic device synchronously loads the compensated historical distance image, the current frame's two-dimensional distance image, and the calibrated fusion weights, with all three sets of data having the same row and column dimensions. The device traverses the image row by row, initiating parallel fusion operations for each row of pixel data read, generating a temporary pixel row array during single-row operations. For a single pixel, the device extracts the pixel distance value of the current frame, the pixel distance value of the compensated historical image, and the fusion weight for that location, with each weight corresponding to the proportion of the two images in the fusion result. The fused pixel value D is then calculated using a weighted average. enhanced =W×D curr +(1-W)×D hist For example, a certain pixel D curr =60m, D hist =58m, after calibration the fusion weight W=0.65, then D enhanced=0.65×60+0.35×58=59.3m, and the calculation result is directly written into the corresponding coordinate position of the enhanced image.

[0128] After all pixel operations in a single row are completed, the temporary pixel row array is concatenated to the enhanced image buffer, and all rows are traversed and operated in sequence. After all pixel calculations are completed, the integrity of the entire image is checked, invalid empty pixels are checked and their status is marked, and finally an enhanced distance image with richer data details, lower noise, and stronger continuity is generated.

[0129] This application allocates information proportions from two images, giving higher weight to high-reliability data and reasonably suppressing low-reliability data. While preserving real-time detection information, it uses historical time-series data to improve image details and suppress random noise, thereby enhancing the overall image performance. The enhanced distance image has uniform pixel data, complete edge features, and preliminary noise suppression, reducing the computational pressure of subsequent spatial edge-preserving filtering. This ensures that the filtering process accurately distinguishes between edge and noise regions, providing input data for filtering and subsequent coordinate inverse projection processes.

[0130] S4. Perform spatial edge-preserving filtering on the enhanced range image, and then back-project the filtered range image into a 3D point cloud through inverse coordinate transformation. After correcting the coordinates, output the preprocessed radar point cloud stream.

[0131] Specifically, the output preprocessed radar point cloud stream includes:

[0132] Spatial edge-preserving filtering is performed on the enhanced distance image according to the partition, and the filtering strategy is dynamically adjusted in combination with the data identifier of the distance image to obtain the filtered distance image;

[0133] By retrieving the projection parameters corresponding to different partitions, and performing inverse coordinate transformation on the filtered distance image partition by partition, the three-dimensional point cloud is restored.

[0134] Data credibility levels are determined based on the reliability of pixel distance values ​​and motion confidence, and coordinate deviation correction is performed on the 3D point cloud according to the data credibility level.

[0135] The corrected 3D point cloud is grouped, cached, and time-ordered, and the preprocessed radar point cloud stream is output.

[0136] Among them, performing inverse coordinate transformation in each partition refers to reusing the projection curvature and angle compensation coefficients of different partitions during forward spherical projection, mapping the two-dimensional pixel coordinates to three-dimensional coordinates according to the preset coordinate transformation relationship, and compensating for the angle offset caused by forward projection.

[0137] Group caching and timing warping refers to storing data in groups according to the receiving timing of the original radar point cloud data stream, removing abnormal point cloud data, aligning the receiving timing within a programmable logic device, and outputting the radar point cloud stream.

[0138] Even after spatiotemporal fusion processing, the enhanced distance image still retains radar random detection noise and environmental electromagnetic interference noise. If directly introduced into the coordinate inverse transformation process, it will be converted into coordinate errors in the 3D point cloud, reducing the accuracy of the point cloud data. Uniform filtering parameters across the entire domain cannot simultaneously achieve noise suppression and edge feature preservation. Different types of pixels have different tolerances to the intensity of filtering smoothing. The original effective ranging pixels carry the true detection information, and excessive smoothing will cause loss of details. Interpolation to complete pixels is prone to generating secondary noise.

[0139] The spatial edge-preserving filtering by partition refers to allocating parallel computing channels to the corresponding partitions of the central main field of view, the edge side field of view, and the blind transition area, and simultaneously carrying out edge-preserving filtering operations within the spatial domain. The spatial edge-preserving filtering is implemented using a bilateral spatial filtering method, relying on the spatial neighborhood distance and pixel numerical gradient to form a dual constraint, distinguishing between flat areas and edge areas of the image, and preserving image contour features while filtering out noise. The data identifiers of the distance image include partition identifiers, empty pixel interpolation identifiers, and effective ranging identifiers. The dynamic adjustment filtering strategy refers to changing three types of operating parameters in real time, namely the filtering window size, smoothing intensity, and edge gradient threshold, according to the identifier type corresponding to the pixel.

[0140] The programmable logic device loads the global pixels and data identifiers of the enhanced distance image in parallel, with the row and column dimensions of the two matrices remaining consistent. Image segmentation is performed based on the image partition identifiers, decomposing the entire frame into three partition sub-images, each connected to three parallel filtering channels, with the number of channels corresponding to the number of partitions. Preset filtering parameters are provided for the three types of partitions: a 5×5 pixel filtering window is configured for the central main field of view, with an edge gradient threshold set to 18; a 7×7 pixel filtering window is configured for the edge side fields of view, with an edge gradient threshold set to 14; and a 9×9 pixel filtering window is configured for the blind transition zone, with an edge gradient threshold set to 10. The size of the partition filtering window increases with the degree of field of view distortion and noise level, while the edge gradient threshold decreases with the noise level, thus adapting to the imaging characteristics of different regions.

[0141] The above filtering parameters are set based on the following principles: the selection of the filtering window size follows the principle that the higher the noise level, the larger the window should be. That is, the noise in the central main field of view is the lowest, so a smaller 5×5 window is used to preserve details to the maximum extent; the noise in the edge side field of view is moderate, so a 7×7 window is used to enhance smoothing ability; the noise in the blind transition area is the strongest, so a 9×9 window is used to fully suppress noise; the selection of the edge gradient judgment threshold follows the principle that the higher the noise level, the lower the threshold should be. That is, in high noise areas, the gradient values ​​between pixels are generally large, so the threshold needs to be reduced to effectively identify the real edges. Therefore, the gradient threshold decreases from 18 in the central main field of view to 10 in the blind transition area; the specific values ​​of the above parameters need to be calibrated according to the statistical data characteristics of the distance image, and those skilled in the art can make adaptive adjustments in the actual process.

[0142] Within a single partition, traversal processing is performed pixel by pixel in row-priority order. During the traversal, pixel filtering parameters are generated and recorded, including pixel row and column coordinates, partition number, identifier type, currently enabled filter window size, edge gradient threshold, and smoothing intensity coefficient. The data identifier is read pixel by pixel and the pixel type is determined. The filtering parameters are dynamically corrected simultaneously. When a pixel is identified as a valid original ranging pixel, the smoothing intensity coefficient is lowered to 0.3 based on the partition's basic parameters to weaken the smoothing effect and retain real detection details. When a pixel is identified as an empty pixel that has been interpolated to complete the pixel, the smoothing intensity coefficient is increased to 0.8 to enhance noise suppression capabilities. When a pixel is identified as an invalid ranging pixel, a smoothing operation is performed using a 9×9 global window and a smoothing intensity coefficient of 1.

[0143] The smoothing intensity coefficient ranges from 0 to 1. A larger value indicates a stronger filtering and smoothing effect. The original effective ranging pixels carry real radar detection information, and excessive smoothing will lead to the loss of edges and details. Therefore, a low value of 0.3 is used to minimize the interference with the original data. The values ​​of the empty pixel interpolation and completion pixels are derived from the neighborhood estimation. They have certain interpolation errors and secondary noise. Stronger smoothing is required to suppress such noise. Therefore, 0.8 is used. Invalid ranging pixels do not contain effective detection information. The maximum value of 1 is used for maximum smoothing and filling. Those skilled in the art can adjust the above smoothing intensity coefficients accordingly in the actual process.

[0144] Spatial domain bilateral filtering defines a filtering window centered on the current pixel, and calculates the spatial domain weight and value domain weight of the pixels within the window. The spatial domain weight w s(i,j) Calculate w using the Gaussian function s(i,j) =exp(-(d i 2 +d j 2 ) / (2σ s 2 )), where d i d j σ represents the row and column distances between neighboring pixels and the current pixel. s The standard deviation is the spatial domain; the range weight is w. v(i) Similarly, w is calculated using the Gaussian function. v(i) =exp(-(I i -I c ) 2 / (2σ r 2 )), where I i For neighboring pixel values, I c σ is the current pixel value. r σ is the standard deviation of the range; s With σ rConfigure the filtering window size according to the partition, for example, the center main field of view σ. s =1、σ r =1 / 3 of the corresponding partition gradient threshold, and so on for other partitions. Multiply the two types of weights to obtain the final fusion weight. Based on the fusion weight, perform a weighted summation on the neighboring pixels within the window to calculate the filtered value of the current pixel. Simultaneously calculate the pixel neighborhood gradient. If the gradient is greater than the partition gradient threshold, it is determined to be an image edge, and only a light weighted operation is performed. If the gradient is less than the partition gradient threshold, it is determined to be a flat region, and a full filtering operation is performed. After all pixels in a single partition are processed, the corresponding partition filtered sub-image is generated and temporarily stored in the partition cache. After all partition operations are completed, the three partition sub-images are integrated according to the original partition stitching rules to generate the filtered distance image.

[0145] This application matches noise and distortion features of different fields of view, effectively filters out various types of noise remaining in the fusion process, and accurately protects the image edges and details corresponding to radar ranging; it blocks the transmission of two-dimensional image noise to three-dimensional point clouds from the source, and improves the purity of subsequent coordinate inverse transformation results.

[0146] In the forward spherical projection stage, projection curvature and angle compensation coefficients are configured for the three field-of-view partitions. This set of parameters is used to compensate for the geometric distortion caused by forward projection, but it will also introduce systematic angle offset errors. If a unified inverse transformation rule is used to complete the restoration of the two-dimensional image to three-dimensional coordinates, it cannot offset the differential systematic deviations in each partition. The final generated three-dimensional point cloud will have problems such as spatial distortion and coordinate offset.

[0147] The projection parameters corresponding to the partitions refer to the projection curvature and angle compensation coefficients configured for each partition in the forward spherical projection stage; performing inverse coordinate transformation partition by partition means independently performing inverse mapping operations on the two-dimensional pixels of each partition according to the image partitioning results; the preset coordinate transformation relationship refers to the inverse mapping operation logic of converting the two-dimensional pixel coordinates corresponding to the forward spherical projection to the three-dimensional radar spherical coordinates, including step-by-step calculation formulas, variable conversion and partition parameter compensation rules; compensating for the angle offset caused by forward projection means substituting the partition angle compensation coefficient in reverse to offset the systematic angle error introduced by forward projection.

[0148] Global variables and fixed parameters are defined. The pixel size of the entire frame of the two-dimensional distance image is 640 columns and 480 rows. The pixel column coordinate u ranges from [0, 639] and the row coordinate v ranges from [0, 479]. The radar radial detection range R is [0.5, 200] meters, the azimuth angle Az is [-π / 2, π / 2] rad, corresponding to an angle range of -90° to +90°, and the elevation angle E1 is [-π / 4, π / 4] rad, corresponding to an angle range of -45° to +45°. The fixed projection parameters of each zone are consistent with the forward projection. For example, the projection curvature of the central main field of view is k1=1, and the angle compensation coefficient Δα1=0.02 rad; the projection curvature of the edge side field of view is k2=1.15, and the angle compensation coefficient Δα2=0.18 rad; the projection curvature of the blind transition zone is k3=1.08, and the angle compensation coefficient Δα3=0.3 rad.

[0149] Forward spherical projection is a transformation operation from three-dimensional radar spherical coordinates to two-dimensional image pixel coordinates. Here, the coordinate transformation relationship is assumed to be its reverse operation. The overall process is divided into pixel coordinate normalization, angle calculation, partition parameter compensation, and boundary determination. Pixel coordinate normalization transforms discrete pixel coordinates to [-1,1] to eliminate the influence of image size, i.e., u norm =2u / (U max -1)-1,v norm =2v / (V max -1)-1,U max =640 is the total number of columns in the image, V max =480 is the total number of rows in the image, u norm v norm Normalized row and column coordinates.

[0150] The original azimuth and elevation angles, i.e., Az, are calculated by combining the zonal projection curvature. raw =k×π / 2×u norm El raw =k×π / 4×v norm Az raw El raw The uncompensated original azimuth and elevation angles are given, where k is the current zone projection curvature; the reverse compensation is used to compensate for the angular offset caused by the forward projection, i.e., Az. out =Az raw -Δα, El out =El raw -Δα, where Δα is the current partition angle compensation coefficient, Az out El out The final output includes azimuth and pitch angles; then, the 3D coordinates are combined and boundary conditions are determined. The pixel distance value R is directly used as the radial distance of the 3D point cloud. The radial distance, final azimuth angle, and final pitch angle are integrated to form the 3D coordinates (R, Az). outEl out If the coordinates exceed the radar's preset detection range, the point cloud is determined to be invalid and is removed.

[0151] The projection curvature and angle compensation coefficients of the three types of partitions are retrieved, and the parameter values ​​are consistent with those of the forward projection. The filtered distance image is then segmented again according to the partitions. After each partition sub-image is bound to the corresponding projection parameters, the inverse coordinate transformation operation is performed. During the operation, the row and column coordinates and pixel distance values ​​of the two-dimensional image are extracted pixel by pixel. The operation is completed in sequence according to the above four-step transformation relationship. At the same time, the mapping relationship between the two-dimensional image coordinates and the three-dimensional coordinates is generated. The record includes the two-dimensional pixel row and column coordinates, the partition to which it belongs, the projection curvature, the angle compensation coefficient, the normalized coordinates, the original angle, the compensated three-dimensional coordinates, and the boundary judgment result.

[0152] A single pixel within the central main field of view is selected as the calculation sample, where the pixel coordinates are u=320, v=240, and the pixel distance is R=50m; the partitioning parameters are k1=1 and Δα1=0.02rad; the normalized calculation is performed on u. norm =2×320÷640-1=0,v norm =2×240÷480-1=0; Original angle solution Az raw =1×π / 2×0=0rad, El raw =1×π / 4×0=0rad; Angle offset compensation Az out =0-0.02=−0.02rad, El out =0-0.02=-0.02rad; The final three-dimensional coordinates (50m,-0.02rad,-0.02rad) of the three-dimensional coordinate combination and boundary determination are within the radar detection range and are determined to be valid point cloud.

[0153] The device presets the radar's full-domain detection space range as the boundary judgment threshold, with radial distance of 0.5m to 200m, azimuth angle of -π / 2 to +π / 2 rad, and elevation angle of -π / 4 to +π / 4 rad. If the calculated three-dimensional coordinates exceed this space range, they are judged as invalid mapped point clouds and are directly discarded. Point clouds with coordinates within the valid range are temporarily stored in the partitioned point cloud cache. After all pixels in a single partition have been processed, a partitioned three-dimensional point cloud is generated. After all partitions have been processed, all valid point clouds from all partitions are integrated to form a full-domain three-dimensional point cloud.

[0154] This application specifically counteracts projection distortion and system angle shift in different zones, effectively improving the coordinate accuracy of restoring two-dimensional images to three-dimensional point clouds; the overall spatial shape of the three-dimensional point cloud conforms to the real radar detection scene, providing a complete computational object for differential coordinate deviation correction.

[0155] The 3D point cloud obtained after inverse coordinate transformation has residual coordinate bias. The magnitude of the bias is directly related to the data quality of the original pixels. The reliability of the pixel distance value is used to characterize the stability of the single-point ranging data, and the motion confidence is used to characterize the reliability of the inter-frame motion estimation and compensation results. The two indicators can comprehensively evaluate the overall quality of a single set of point clouds. If a uniform correction amount is used for all point clouds, high-confidence point clouds will lose the original detection accuracy due to over-correction, while low-confidence point clouds cannot completely eliminate coordinate bias.

[0156] The reliability of pixel distance values ​​and motion confidence are defined as described above. After normalization, the value range of the two indicators is unified as [0,1]. Data reliability level refers to the point cloud quality level divided by the comprehensive score of the two indicators. Coordinate deviation correction refers to the calculation process of compensating and correcting the coordinate error remaining in the inverse transformation.

[0157] The reliability and motion confidence of the pixel distance values ​​associated with each group of 3D point clouds are retrieved. The arithmetic mean of the two indicators is taken as the comprehensive score to classify the data reliability level. A comprehensive score greater than or equal to 0.7 is judged as high reliability; a comprehensive score greater than or equal to 0.4 and less than 0.7 is judged as medium reliability; and a comprehensive score less than 0.4 is judged as low reliability. After traversing all point clouds to complete the level determination, the data reliability level is generated. Simultaneously, the following information is recorded: 3D spatial coordinates, reliability of pixel distance values, motion confidence, comprehensive score, reliability level, coordinates before correction, and coordinates after correction.

[0158] Differentiated coordinate correction amounts and logics are configured for the three data confidence levels. High-confidence level point clouds have extremely small residual deviations, so a fixed small correction amount is used, with a single-dimensional coordinate correction range of ±0.03m, correcting only minor system-level errors and preserving the original data accuracy to the maximum extent. Medium-confidence level point clouds have moderate deviations, so a standard correction amount is used, with a single-dimensional coordinate correction range of ±0.1m, completing routine deviation compensation. Low-confidence level point clouds have strong coordinate uncertainty; in addition to the basic single-dimensional ±0.18m correction, high-confidence point clouds within the eight neighboring regions of the low-confidence level point cloud are searched, and linked correction is performed based on the effective coordinates of the neighboring regions to suppress single-point coordinate distortion. The coordinate correction calculation for the 3D point clouds is completed group by group, with coordinate data before and after correction temporarily cached during the correction process. After all point cloud calculations are completed, the corrected 3D point cloud is obtained.

[0159] The point clouds with high confidence levels (overall score ≥ 0.7) have high reliability of pixel distance values ​​and motion confidence. The residual deviation after coordinate inverse transformation mainly comes from the quantization discretization error of spherical projection, so the correction range is set to ±0.03m. Point clouds with medium confidence levels (0.4 ≤ overall score < 0.7) have additional ranging fluctuations and motion compensation residuals in addition to quantization errors, so the correction range is set to ±0.1m. Point clouds with low confidence levels (overall score < 0.4) have poor data quality and the greatest coordinate uncertainty, so the correction range is set to ±0.18m. The spatial constraints of eight neighboring high-confidence point clouds are introduced for linkage correction to suppress coordinate distortion of isolated low-quality points by utilizing reliable surrounding data. The correction range is determined based on the typical statistical distribution of spherical projection quantization error and motion compensation residual. Those skilled in the art can calibrate and adjust it according to the ranging accuracy index and motion compensation accuracy index of the radar system. The linkage correction specifically calculates the weighted average of the three-dimensional coordinates of all high-confidence point clouds in the eight neighboring areas as the reference coordinates, with the weight being the comprehensive score of each neighboring point cloud, and appropriately corrects the low-confidence point cloud coordinates towards the reference coordinates.

[0160] This application objectively distinguishes differences in point cloud quality, accurately matches the deviation characteristics of point clouds of different levels, and effectively corrects coordinate distortion of low-reliability point clouds while protecting the accuracy of high-reliability original data, thus comprehensively improving the spatial coordinate accuracy of global 3D point clouds.

[0161] The programmable logic device (PLD) employs a multi-channel parallel processing mode. This operation mode can cause misalignment between the timing of the generated 3D point cloud and the original sampling timing of the radar front end. Furthermore, the corrected point cloud set still contains a small number of abnormal point clouds with out-of-bounds coordinates and conflicting parameters. Direct output of these abnormal point clouds would interfere with the normal operation of the backend radar application module. Grouping and buffering refers to grouping point cloud data according to the original radar sampling batches and writing them into the on-chip grouping buffer unit of the PLD. Timing regularization refers to aligning the data according to the original radar sampling timing after removing abnormal point clouds. The original radar point cloud data stream reception timing refers to the native timing rules of radar antenna sampling, echo analysis, and point cloud output.

[0162] The point cloud is grouped according to the radar's original sampling rules, with each group containing 64 continuously sampled point clouds. The setting of 64 point clouds per group is consistent with the number of sampling points per scan line of the radar front end to ensure that the group boundary is aligned with the radar's physical sampling cycle. During the grouping process, a point cloud grouping sequence is generated, including the group number, the original sequence number, the number of point clouds in the group, the coordinate range of the point clouds in the group, and the number of outliers. The on-chip grouping cache unit of the device has a preset storage depth of 16 groups and adopts a first-in-first-out read-write rule. Each group of point clouds is written to the corresponding cache location in sequence according to the group number. The cache depth of 16 groups is consistent with the frame storage depth of the aforementioned circular frame buffer to ensure the matching of data flow. Those skilled in the art can adjust the cache depth according to the radar's sampling rate and the on-chip storage capacity of the programmable logic device.

[0163] During the cache reading phase, abnormal point clouds are screened group by group, using the radar's full-domain detection range as the criterion. Point clouds with coordinates exceeding the range of radial distance (0.5m~200m), azimuth (-π / 2~+π / 2 rad), and elevation (-π / 4~+π / 4 rad) are identified as coordinate anomalies and directly removed. Point clouds with negative radial distances are identified as parameter anomalies and removed simultaneously. After anomaly cleanup, all groups in the cache are time-series rearranged according to the original time sequence number to correct time sequence misalignment caused by parallel processing, integrating discrete group data into a continuous data stream. After time sequence alignment is completed, data is continuously output in a pipeline manner, ultimately generating a standardized preprocessed radar point cloud stream.

[0164] This application adapts to the on-chip storage architecture of programmable logic devices, effectively manages the flow of large-capacity point cloud data, avoids data flow blockage, and improves the overall quality of point clouds; ensures that the output point cloud stream is consistent with the original radar sampling timing, and seamlessly connects to downstream modules such as target detection, imaging, and data storage; the output standardized radar point cloud stream can be directly applied to various radar business scenarios.

[0165] Example 2

[0166] like Figure 3 As shown, this embodiment provides a high-speed radar point cloud preprocessing system based on programmable logic, including a mapping module, an estimation module, a fusion module, and an output module.

[0167] The mapping module is used to map the real-time received radar point cloud data stream into a two-dimensional distance image of the current frame through spherical projection within a programmable logic device, and store it in a circular frame buffer.

[0168] The estimation module is used to select historical distance images from the circular frame buffer that have a similarity to the two-dimensional distance image of the current frame that meets the preset conditions, construct image pyramids respectively, and perform inter-frame motion estimation layer by layer to generate motion fields and their motion confidence.

[0169] The fusion module performs motion compensation mapping on the historical distance image based on the motion field to obtain the compensated historical distance image. The compensated historical distance image is then fused with the two-dimensional distance image of the current frame pixel by pixel in a spatiotemporal manner. The fusion weights are adjusted according to the motion confidence and the reliability of the pixel distance values ​​to generate an enhanced distance image.

[0170] The output module is used to perform spatial edge-preserving filtering on the enhanced range image, and then back-project the filtered range image into a 3D point cloud through inverse coordinate transformation. After correcting the coordinates, the preprocessed radar point cloud stream is output.

[0171] The above description is merely a preferred embodiment of this application. The scope of protection of this application is not limited to the above embodiments. All technical solutions falling within the scope of this application's concept are within the scope of protection of this application. It should be noted that for those skilled in the art, any improvements and modifications made without departing from the principles of this application should also be considered within the scope of protection of this application.

Claims

1. A high-speed preprocessing method for radar point clouds based on programmable logic, characterized in that, include: Within the programmable logic device, the real-time received radar point cloud data stream is mapped online through spherical projection into a two-dimensional distance image of the current frame and stored in a circular frame buffer. From the circular frame buffer, select historical distance images whose two-dimensional distance image similarity to the current frame meets the preset conditions, construct image pyramids respectively, and perform inter-frame motion estimation layer by layer to generate motion fields and their motion confidence. Motion compensation mapping is performed on historical distance images based on the motion field to obtain compensated historical distance images. The compensated historical distance images are then spatiotemporally fused pixel by pixel with the two-dimensional distance images of the current frame. The fusion weights are adjusted according to the motion confidence and the reliability of the pixel distance values ​​to generate enhanced distance images. Spatial edge-preserving filtering is performed on the enhanced range image. The filtered range image is then back-projected into a 3D point cloud through inverse coordinate transformation. After coordinate correction, the preprocessed radar point cloud stream is output.

2. The high-speed preprocessing method for radar point clouds based on programmable logic according to claim 1, characterized in that, The online mapping to the two-dimensional distance image of the current frame includes: The radar point cloud data is divided into zones according to the radar detection field of view, and the effective point cloud data is filtered out. The spherical projection parameters are dynamically corrected frame by frame. The corresponding projection parameters are configured according to the partition. The effective point cloud data is mapped to two-dimensional pixel coordinates in parallel partitions to obtain the two-dimensional distance image of the current frame. The system identifies empty pixels formed by the sparse point cloud within a 2D distance image, performs interpolation processing based on the pixel distribution density of the empty pixel's neighborhood, adds data identifiers to the 2D distance image, and then stores it in a circular frame buffer.

3. The high-speed preprocessing method for radar point clouds based on programmable logic according to claim 2, characterized in that, The partition includes the central main field of view, the edge side fields of view, and the blind transition zone; The frame-by-frame dynamic correction of spherical projection parameters involves configuring corresponding projection parameters according to the partitions. Specifically, based on the pixel distribution density and angle mapping deviation value of the previous frame's two-dimensional distance image, the spherical projection parameters are iteratively updated frame by frame, and projection curvature and angle compensation coefficients are configured for the three partitions respectively based on the spherical projection parameters. The pixel distance values ​​of the two-dimensional distance image represent the radial distance of the point cloud; when multiple valid point cloud data are mapped to the same pixel, they are combined with echo intensity and effective ranging identifier for hierarchical weighted fusion.

4. The high-speed preprocessing method for radar point clouds based on programmable logic according to claim 3, characterized in that, The generated sports field and its motion confidence level include: Construct an image pyramid by dividing the selected historical distance image and the current frame's two-dimensional distance image into regions; Inter-frame motion estimation is performed layer by layer from the top to the bottom of the pyramid. The displacement vector of the adjacent upper layer is used as the prior offset for the current lower layer calculation, and the motion field is generated by iterative correction. By combining matching residuals, regional motion consistency, effective ranging identifiers, and fluctuations in pixel distance values, motion confidence is calculated pixel by pixel, and the motion confidence is corrected by combining image region attributes.

5. The high-speed preprocessing method for radar point clouds based on programmable logic according to claim 4, characterized in that, The aforementioned regional image pyramid construction refers to dividing the image region into a central main field of view, an edge side field of view, and a blind transition zone, and assigning a downsampling factor to different image regions; The layer-by-layer inter-frame motion estimation refers to performing block matching in the top-level image of the pyramid according to image regions to generate an initial motion field, narrowing the pixel matching search range from the top to the bottom of the pyramid, and performing neighborhood smoothing correction. The pixel-by-pixel motion confidence is calculated by weighting and fusing the fluctuation range of the matching residual, regional motion consistency, effective ranging identifier, and pixel distance value, and then correcting the motion confidence by configuring correction coefficients according to the image region attributes; the regional motion consistency characterizes the dispersion of the neighborhood displacement vector.

6. The high-speed preprocessing method for radar point clouds based on programmable logic according to claim 5, characterized in that, The generated enhanced distance image includes: Based on the pixel-by-pixel displacement vector corresponding to the motion field, motion compensation mapping is performed on the historical distance image to obtain the compensated historical distance image; The fusion weights are adjusted by combining motion confidence and pixel distance values, and the fusion weights are calibrated by region by combining image region attributes. Based on the calibrated fusion weights, the compensated historical distance image and the current frame's two-dimensional distance image are spatiotemporally fused pixel by pixel to generate an enhanced distance image.

7. The high-speed preprocessing method for radar point clouds based on programmable logic according to claim 6, characterized in that, The adjustment of the fusion weights is to characterize the reliability of inter-frame motion matching with motion confidence. The reliability of pixel distance values ​​is obtained by integrating the effective ranging identifier and the fluctuation range of pixel distance values. After normalizing the motion confidence and the reliability of pixel distance values, the fusion weights are generated pixel by pixel. The regional calibration is to adjust the ratio of fusion weights within the corresponding image region according to the image region attributes.

8. The high-speed preprocessing method for radar point clouds based on programmable logic according to claim 7, characterized in that, The output preprocessed radar point cloud stream includes: Spatial edge-preserving filtering is performed on the enhanced distance image according to the partition, and the filtering strategy is dynamically adjusted in combination with the data identifier of the distance image to obtain the filtered distance image; By retrieving the projection parameters corresponding to different partitions, and performing inverse coordinate transformation on the filtered distance image partition by partition, the three-dimensional point cloud is restored. Data credibility levels are determined based on the reliability of pixel distance values ​​and motion confidence, and coordinate deviation correction is performed on the 3D point cloud according to the data credibility level. The corrected 3D point cloud is grouped, cached, and time-ordered, and the preprocessed radar point cloud stream is output.

9. The high-speed preprocessing method for radar point clouds based on programmable logic according to claim 8, characterized in that, The step-by-step coordinate inverse transformation refers to reusing the projection curvature and angle compensation coefficients of different partitions during forward spherical projection, mapping the two-dimensional pixel coordinates to three-dimensional coordinates according to a preset coordinate transformation relationship, and compensating for the angle offset caused by forward projection. The grouping and timing normalization refers to storing data in groups according to the receiving timing of the original radar point cloud data stream, removing abnormal point cloud data, aligning the receiving timing within the programmable logic device, and outputting the radar point cloud stream.

10. A high-speed radar point cloud preprocessing system based on programmable logic, used to implement the high-speed radar point cloud preprocessing method based on programmable logic as described in any one of claims 1-9, characterized in that, include: The module consists of a mapping module, an estimation module, a fusion module, and an output module. The mapping module is used to map the real-time received radar point cloud data stream into a two-dimensional distance image of the current frame through spherical projection within a programmable logic device, and store it in a circular frame buffer. The estimation module is used to select historical distance images from the circular frame buffer that have a similarity to the two-dimensional distance image of the current frame that meets the preset conditions, construct image pyramids respectively, and perform inter-frame motion estimation layer by layer to generate motion fields and their motion confidence. The fusion module performs motion compensation mapping on the historical distance image based on the motion field to obtain the compensated historical distance image. The compensated historical distance image is then fused with the two-dimensional distance image of the current frame pixel by pixel in a spatiotemporal manner. The fusion weights are adjusted according to the motion confidence and the reliability of the pixel distance values ​​to generate an enhanced distance image. The output module is used to perform spatial edge-preserving filtering on the enhanced range image, and then back-project the filtered range image into a 3D point cloud through inverse coordinate transformation. After correcting the coordinates, the preprocessed radar point cloud stream is output.

Images